Start of Lecture: February 10, Chapter 6: Scheduling

Transcription

1 Start of Lecture: February 10,

2 Reminders Exercise 2 due this Wednesday before class Any questions or comments? 2

3 Scheduling so far First-Come-First Serve FIFO scheduling in queue without preempting processes Round-Robin Scheduling FCFS with preempting and same time quantum for each process Shortest Job First choose shortest jobs first showed that this gives the minimal average waiting time long jobs might starve 3

4 Stefan and Jon s suggestion: alternate longest and shortest jobs Shortest job first: possible starvation for longest jobs Alternating longest and shortest jobs pros no longer starves the longest processes could still have convoy for some periods of time, but since alternating and not taking strict order, might end up re-ordering and jumping stuck process ahead of the long processes Alternating longest and shortest jobs cons without preempting, can still have convoy issues medium-length process might starve 4

5 Predicting CPU-burst lengths for SJF Can only estimate the length hope that it might be similar to (most recent) previous lengths One approach: use average of all previous bursts how estimate average without storing all previous values? m t = 1 t tx i=1 a i! m t = 1 t (sum t 1 + a t )! OR m t = (t 1)m t 1 + a t t OR m t = m t 1 + a t m t 1 t any disadvantages to this approach? Another approach: use exponential averaging 5

6 Exponential averaging for CPU bursts t n = actual length of n th CPU burst n+1 = predicted length of next CPU burst n+1 = t n +(1 ) n where 0 apple apple 1 and commonly =0.5 τ i t i time CPU burst (t i ) "guess" (τ i )

7 Research topic: use machine learning Exponential averages do not use properties of the process to predict what might come next Collect data on process, such as length of associated code, number of open files, number children, total time completed so far, etc. AND its corresponding CPU-burst CPU-burst is the variable to predict Now can use (statistical) optimization techniques to learn a function mapping the detailed information about the process to its most likely next CPU-burst length Info about process X w = y learned mapping CPU-burst length 7

8 Example of Shortest-remaining-time-first Now we add the concepts of varying arrival times and preemption to the analysis ProcessAarri Arrival TimeT Burst Time P P P P Preemptive SJF Gantt Chart P 1 P 2 P 4 P 1 P Average waiting time = [(10-1)+(1-1)+(17-2)+5-3)]/4 = 26/4 = 6.5 msec 8

9 Priority Scheduling A priority number is associated with each process Have to decide if low numbers mean high priority (true in many systems) or if low numbers mean low priority FCFS and SJF are special instances of priority scheduling FCFS assigns equal priority to each process (so just removed from the queue in order that arrived) SJF assigns higher priority to short process: priority = inverse of predicted next CPU burst length 9

10 Example of Priority Scheduling Assume all processes arrive at same time: ProcessA arri Burst TimeT Priority P P P P P Priority scheduling Gantt Chart P 21 P 52 P 1 P 3 P Average waiting time = 8.2 msec 10

11 Implementation of priority scheduling Keep linked list and scan the entire list for priority What is the worst case removal runtime? O(n) What is the runtime for adding an element? Keep a sorted linked list O(1) What is the worst case removal runtime? O(1) What is the runtime for adding an element? O(n) Use a priority queue: process with highest priority is at front of queue, but list not necessarily in order 11

12 Priority queue (binary heap) If low number corresponds to a higher priority, then use a min heap; if high number to high priority, the max heap Inserting a new process into heap: O(log n) Removing a process from heap: O(log n) Using binary heaps as priority queue efficient! Min Heap Max Heap 12

13 Inserting into and deleting from heap Check out the applet here for a max heap demo: When remove the head of the queue, that returns the highest priority process When insert, it first adds the new process with priority to leaf node, then does a re-heapify where it swaps nodes until heap property satisfied again 13

14 Advantages of priority scheduling Priority scheduling encapsulates the other scheduling algorithms we ve discussed in one nice unified way: by assigning priorities depending algorithm properties such as CPU-burst lengths Flexibility without incurring much overhead, due to efficiency of priority queues Natural way to order processes, rather than more arbitrary, fixed heuristics like first one to arrive 14

15 Disadvantage of priority scheduling Priority scheduling can result in in-definite blocking or starvation Particularly in a heavily-loaded system, where new processes are constantly created, low-priority process might wait forever until system is finally lightly-loaded or until it crashes and loses them One solution: age the priority of low-priority processes, so that priority increases with wait time 15

16 How do you select priorities? Priorities can be defined externally set by criteria outside the operating system importance of process type and amount of funds being paid for computer use Priorities can be defined internally define measurable quantities to compute priority. Examples that been used: time limits memory requirements number of open files ratio of average I/O burst to average CPU-burst 16

17 Comparison of scheduling policies Unfortunately, performance varies substantially depending on characteristics of jobs entering system (the job mix): definitive comparisons impractical we will come back to more experimental ways to evaluate a scheduling policy for your specific system Some rules of thumb for now: FCFS performs better for long processes and favours CPU-bound jobs FCFS not suitable for interactive jobs; RR not suitable for batch jobs SJF is risky, since long processes may suffer CPU starvation Don t forget processing times: overhead of FCFS negligible, moderate in RR and higher in SJF 17

18 Video Break: brought to you by another fantastic classmate 18

19 Exercise: Some scheduling thought questions (that I want you to think about) How deal with different types of processes on same system that could be scheduled better with different algorithms? multi-level queue scheduling, where separate queues for different processes with different scheduling algorithms on each queue, with one higher-level scheduler determining which of the queues gets to give a process right now What are some heuristics you might use for a real scheduler? For RR, ~80 percent of CPU bursts should be shorter than time quantum, to minimize context-switches whilst also avoiding degenerating to FCFS Give higher priority to currently running process to avoid context-switching 19

20 How deal with different types of processes? There is no one-glove-fits-all solution for scheduling; other schemes developed for handling a mixed collection of jobs (e.g. batch, interactive, CPU-bound) Multi-level queue scheduling have separate queues for different process types Multi-level feedback queue scheduling a more flexible variant of multilevel queue scheduling 20

21 Multi-level Queue Scheduling Separate ready queue for each process class with possibly different scheduling algorithms for each queue highest priority job classes system processes priority-based primary CPU scheduling interactive processes Round-Robin CPU lowest priority batch processes FCFS 21

22 Multi-level Queue Scheduling Properties Natural separation between foreground and background jobs, since have separate queues Can take advantage of simple scheduling algorithms, since know what types of processes will be in each queue Can even specify percentage CPU time given to what process types e.g. 80 percent to queue 1, 20 to queue 2 Disadvantages some overhead to maintain different queues (which might possibly be offset by the fact that each queue is now shorter) process type fixed but what if a process has varied properties? 22

23 Multi-level feedback queue scheduling Allows processes to move between queues, depending on processes current properties One example: using CPU-burst length to determine type If took longer than 2 milliseconds, moved to medium priority queue If preempted after 8 milliseconds, moved to low, FCFS queue High priority: short interactive processes this scheme. Quantum = 2 Quantum = 8 primary CPU scheduling CPU Medium priority: waits until no high priority processes FCFS Low priority: waits until other queues empty 23

24 General multi-level feedback queue scheduler Defined by following parameters: Number of queues Scheduling algorithm for each queue Method used to determine when to upgrade process to a higher level (higher priority) queue e.g. waiting/starving in lower queue, so increase priority Method used to determine when to demote a process to a lower level (lower priority) queue previous example was based on CPU-burst length; others possible Method to initially assign a process to a queue 24

25 Multi-Processor Scheduling What do you think are some issues for scheduling on multiple processors? Should they share a ready queue? Have their own queues? Do multiple processors versus multicore processors have to be treated differently? Load balancing between processors processes migrate between processors to keep workload evenly distributed Should you use pull migration (idle CPU pulls a task from elsewhere) or push migration (specific task periodically checks load and moves processes) Processor affinity process wants to stay on same processor because recent variables all stored in cache; has to repopulate on new processor 25